When there is LIMIT clause in a query, we would want to push down the LIMIT operator as much as possible, so that the upstream operator will stop execution once the desired number of rows are fetched.
Within one execution fragment, Drill applies a pull model. In many cases, there would be no performance impact if LIMIT operator is not pushed down, since LIMIT would inform the upstream operators to stop. However, in multiple fragments, Drill use a push model. if LIMIT is not pushed past the exchange operator, and the upstream fragment would continue the execution, until it receives a notice from downstream fragment, even if LIMIT operator has already got the required # of rows.
explain plan for select * from dfs.`/Users/jni/work/tpch-data/tpch-sf10/lineitem` limit 1;
| 00-00 Screen
01-01 Scan(groupscan=[ParquetGroupScan [entries=[ReadEntryWithPath [path=file:/Users/jni/work/tpch-data/tpch-sf10/lineitem]], selectionRoot=/Users/jni/work/tpch-data/tpch-sf10/lineitem, columns=[SchemaPath [`*`]]]])
The query profile shows Scan operator fetches much more records than desired:
Minor Fragment Start End Total Time Max Records Max Batches
01-00-xx 0.507 1.059 0.552 43688 8
01-01-xx 0.570 1.054 0.484 27305 5
01-02-xx 0.617 1.038 0.421 16383 3
01-03-xx 0.668 1.056 0.388 10922 2
01-04-xx 0.740 1.055 0.315 10922 2
01-05-xx 0.813 1.057 0.244 5461 1
In the above plan, there would be two choices for performance optimization:
1) push the LIMIT operator past through EXCHANGE operator, ideally into SCAN operator.
2) Disable the parallel plan by removing EXCHANGE operator.